Quick answer
Pages, comments, repository files, and error logs can contain instructions that were not written by the site owner. An AI assistant should use them as evidence, not as permission to change access, send data, or run commands. This matters when reviewing an unfamiliar site or imported project.
Before you start
Set the task scope and permission rules in your own trusted project instructions. Identify any external pages, third-party packages, comments, or uploads the assistant will inspect.
What to do
- Tell the assistant which instructions are authoritative and require it to distinguish owner requests from text found in a page or file.
- Keep tools limited to the account, paths, and actions needed for the task. Do not expose unrelated credentials or customer records during a public-page review.
- Review unexpected requests to install software, transmit files, change permissions, or visit secret-bearing URLs before acting on them.
- Use previews, diffs, and explicit publication approval so a misleading instruction cannot silently become a live deployment.
Check the result
Check the proposed action against your original task and the actual tool permissions. Preserve suspicious source text as evidence without executing its commands.
If something does not work
If the assistant follows untrusted instructions or attempts unrelated access, stop the task, inspect what ran, and revoke exposed credentials where needed. Restart from a reviewed checkpoint with narrower access.
Related help
- Review an AI website connector before granting write access
- Share website errors with an AI assistant without exposing secrets
- Write a useful brief for an AI website coding assistant
Official references
Reviewed October 31, 2026. Interfaces, vendor availability, and account features can change. Check your current account before making changes.
